Monitoring precursory volcanic activity: Applying convolutional neural networks to the decades-long ASTER archive

The next generation of thermal infrared (TIR) sensors will provide higher spatial and temporal resolution data than currently available. These include the ISRO-CNES’s Thermal infraRed Imaging Satellite for High-resolution Natural Resource Assessment (TRISHNA), ESA’s Land Surface Temperature Monitoring (LSTM), and NASA-ASI’s Surface Biology and Geology (SBG) missions. The near-daily coverage at ~60m spatial resolution will be invaluable for volcano monitoring but introduces new challenges. The large and complex data volumes from these missions require new advanced analytical approaches for effective detection of volcanic unrest. The 25-year archive of 90 m spatial resolution TIR data from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) has accurately detected both large surface temperature variations during eruptive activity and subtle anomalies (1-2K) associated with degassing and precursory summit activity. Preliminary studies on eruption forecasting potential used ASTER data to constrain models of magmatic and geothermal processes, both crucial for improving hazard mitigation. A machine learning (ML) version of the Automated Spatiotemporal Thermal Anomaly Detection (ASTAD) algorithm, a CNN-based model specifically designed for ASTER data, achieved improved detection rates. CNN models are well suited for extracting spatial and thermal features as well as identifying subtle anomalies. The combination of ASTER’s spatial resolution and ASTAD-ML’s pattern recognition capabilities allows us to retrospectively test the approach globally in preparation for future missions. Here, we show the capability of ASTAD-ML by designing a global cloud-based AI platform populated with ASTER data. We applied the ASTAD-ML model to 100 representative volcanoes spanning a wide range of thermal, morphological, and volcanological activity types. The model includes both day and night data, as well as scenes typically discarded due to cloud cover or partial data loss/stripping. We evaluated both pixel-based and event-based performance, achieving BF1 and F1 high scores of 0.80 and 0.89, respectively. The ASTAD-ML model's pattern recognition capabilities both expanded the usable dataset and improved the accuracy of automatic early volcanic unrest detection. The methodology is highly adaptive, and further testing is ongoing in preparation for these future high spatial resolution TIR sensors, enabling significantly improved monitoring of global volcanic activity.